Better datetime axis compatibility

[nvaytet]

Computing limits on datetime axes is a little funny with data has that passed through matplotlib.
Datetimes are converted to floats for pcolormesh coordinates (2d images) but not for lines.

This PR tries to handle the different cases in a more consistent way, which allows us to plot 1d and 2d data on the same axes with datetime coords (which failed previously). It is not a common use case but was requested by users.

Example:

import plopp as pp
import numpy as np
import scipp as sc

# 2d data
t = np.arange(
    np.datetime64('2017-03-16T20:58:17'), np.datetime64('2017-03-16T21:15:17'), 20
)
time = sc.array(dims=['time'], values=t)
z = sc.arange('z', 50.0, unit='m')
v = np.random.random(z.shape + time.shape)
da = sc.DataArray(
    data=sc.array(dims=['z', 'time'], values=10 * v, variances=v),
    coords={'time': time, 'z': z},
)
f = pp.plot(da)

# 1d data
v = np.random.rand(time.sizes['time'])
a = sc.DataArray(
    data=sc.array(dims=['time'], values=10 * v, variances=v),
    coords={'time': time},
)
a.plot(ax=f.ax, color='r')

[jl-wynen]

This is in the common backend module. Shouldn't it be in the matplotlib module?

[nvaytet]

Hmm, good point.

That means we need a different bounding box implementation for each backend

We can do it with a common bounding box, but it just means we have to convert all the values in the coordinate to floats instead of converting only the min and max. The performance penalty is probably minimal, considering all the other things we do for every update.

[jl-wynen]

Can you add a test?